Theme

Scaffolding, not weights

Explore this theme’s evidence over time →

Sources

Attribution and reporting linked by this record; links do not imply independent verification.

Welcome to December 12, 2025 (theinnermostloop.substack.com)

Progress relocates from the model to the harness around it: framing a task well, or wrapping a model in tools, buys more than the next checkpoint. If it holds, superintelligence arrives as unlocked existing weights.

Development timeline

Newest reporting dates first. Reporting dates may differ from event dates.

2026-08-23: An agent architecture lifts a model from 30% to a perfect 100

2026-08-13: Stock tooling turns 30% into 96.2% for about $540

2026-08-06: A vendor-neutral spec packs skills and servers into portable bundles

2026-07-30: Two API settings triple a benchmark score on six times fewer tokens

2026-07-26: A lab deletes 80% of its agent's system prompt

2026-07-21: A model trained only on short tasks solves tasks 32 times longer

2026-07-17: A harness hits 99% on a benchmark with zero weight changes

2026-07-08: A skill ships for meta-harnesses that rewrite their own scaffolding

2026-07-05: A rival credits scale, long-horizon RL and a constitution

2026-07-02: A system ports learned behaviors to fresh model bases as fast as they ship

2026-06-28: A researcher says the plateau is far out with enough scaffolding

2026-06-25: An agent mines its weaknesses and rewrites its own scaffolding

2026-06-24: A command turns any document into a reusable skill

2026-06-21: A robotics lead declares vision-language-action models dead

2026-06-19: An agent learns a chore by watching a recorded demonstration

2026-06-09: A deterministic retrieval layer takes viral accuracy from 17% to over 90%

2026-05-26: A skill document becomes the trainable state of a frozen agent

2026-05-26: Researchers argue language models need sleep

2026-05-25: An IDE startup is elevated into a frontier lab

2026-05-20: A stack ties co-scientist, evolution and notebooks into one science platform

2026-05-16: A compact memory state is grafted onto a frozen model

2026-05-15: A meta-system builds its own harnesses and sets a coding record

2026-05-13: Users metaprompt an agent to write its own objective file

2026-05-11: A lab winds down its fine-tuning API

2026-05-11: A self-writing harness takes the top spot in global token rankings

2026-05-09: A Fields medallist gets PhD-level research in about an hour

2026-05-06: Driving a screen costs forty-five times more than calling an API

2026-03-31: A research loop writes the strategies for its own outer loop

2026-03-31: Agent control logic becomes a portable natural-language object

2026-03-27: A scaffolding SDK claims 36% where a frontier model reached 0.25%

2026-03-16: The first open-source agentic AI physicist launches

2026-02-26: One product routes tasks across nineteen models in parallel

2026-01-27: A coding agent rewrites its own codebase every day

2026-01-25: Claude Code gets a project management cortex

2026-01-24: Coding agents scale best in a corporate hierarchy

2026-01-24: Claude Code gets tool search as Grokipedia traffic rises 100x

2026-01-11: Repeating a prompt twice improves performance for free

2026-01-05: Windows makes agents first-class apps

2026-01-04: An image model is used as persistent visual memory

2026-01-04: Recursive models handle contexts 100x their window

2026-01-02: A model takes over its own context via a Python REPL

2025-12-29: A System 3 outer loop cuts reasoning steps 80%

2025-12-28: Epoch finds 15% on SWE-bench Verified from prompt restructuring alone

2025-12-27: Adobe extracts large causal models from LLMs

2025-12-24: A harness takes ARC-AGI-2 to 75% without touching the model

2025-12-23: Claude Code adds Language Server Protocol support

2025-12-22: A Meta-RL framework lets agents learn from feedback without gradient updates

2025-12-22: Updating query projections beats generating thinking tokens on long contexts

2025-12-19: Google works to make retrieval differentiable in pre-training

2025-12-16: Engineers describe the job as prompting and sanity-checking

2025-12-16: NVIDIA open-sources Nemotron 3 with data and RL environments

2025-12-13: Terence Tao solves Erdős #1026 with machine assistance

2025-12-13: Thinking Machines releases Tinker for fine-tuning

2025-12-13: Zoom's federated swarm beats the monolith on Humanity's Last Exam

2025-12-12: Google commoditizes autonomous investigation via API

2025-12-12: Kilpatrick: superintelligence arrives as unlocked weights

2025-12-12: Artificial Analysis releases the Stirrup harness

Knowledge graphopen in browser →

Referenced by

The harness writes itself

Cite or correct this record

Citation formatCitation text

Scaffolding, not weights
Innermost Loop KB — secondary record
Reporting date: unknown
KB permalink: https://nicholsn.github.io/innermost-loop-kb/themes/scaffolding-over-weights
Attributed sources:
- Welcome to December 12, 2025 — https://theinnermostloop.substack.com/p/welcome-to-december-12-2025
This citation does not imply independent verification of linked-source claims.

Suggest a correction — opens a GitHub draft for you to review and submit.